Kimi K3's ~3T parameters spread across 59 Cerebras WSE-3T wafers, weights stay resident
petrusenko_max · x · 2026-10-26
Cerebras detailed how a 3T-parameter Kimi K3 model maps onto 59 WSE-3T wafer-scale chips: two linear-attention layers per wafer (35 wafers) and one dense layer per wafer (24 wafers), with 4-bit weights and remaining SRAM holding an 8-bit KV cache. The key design choice: weights stay resident on-wafer, and only activations cross between wafers.
More from Infra
- Lenovo expands enterprise AI Library: 1-week deployment, up to 70% lower token cost — shashib · 2026-10-26
- Fireworks valuation soars from $4B toward $40B as inference clouds bet on falling prices — FinanceYF5 · 2026-10-26
- Temporal raises $550M at $12.5B valuation, up from $5B seven months ago — FinanceYF5 · 2026-10-26
- Bindu Reddy: GPU arbitrage isn't a technology or a defensible business — bindureddy · 2026-10-26
- Can a dual-Xeon HP Z8 with 768GB RAM and one RTX 3090 still serve local LLMs? — libbyt91 · 2026-10-26
- S&P Global: Big Tech to spend over $7 trillion on data centers and AI capex by 2030 — Beth_Kindig · 2026-10-26